Project summary

I work on rodents, small-mammal conservation and animal welfare in Nigeria. One thing I have become increasingly interested in is how people will use AI when they need practical advice about animals.

Someone can now ask an AI system what to do about rats in a poultry farm, rodents in a house, animals suspected of carrying disease, or how to reduce crop losses. The answer may influence what that person actually does. However, I have not seen much work testing whether AI gives good advice in these situations from both a scientific and animal-welfare perspective.

I would like to investigate this using rodent management as a case study.

Rodent control presents a genuine challenge. Rodents can damage food and crops and can be involved in disease transmission, so there are legitimate reasons for controlling their numbers. At the same time, some commonly used methods can cause significant suffering, affect non-target animals, or create other environmental and public-health problems.

The project will develop a set of realistic questions about rodent management and use them to test several AI systems. The questions will be based on situations that people actually encounter, including household rodent problems, agricultural settings, disease-risk situations, trapping, rodenticides and alternatives such as fertility control.

We will then have people with relevant expertise assess the answers. We will look at whether the AI gives biologically correct information, whether it considers animal welfare, whether it identifies risks to people and non-target animals, and whether the advice is actually appropriate for the situation described.

I am particularly interested in finding out where AI gets things wrong. For example, does it automatically recommend poison when less harmful options are available? Does it understand that different rodent species and situations require different approaches? Does it distinguish between controlling rodents because of a genuine disease risk and simply killing animals because they are considered pests?

The results could provide a useful starting point for improving the way animal welfare is considered in AI systems.

What are this project's goals? How will you achieve them?

The first goal is to develop a collection of realistic rodent-management questions that can be used to test AI systems.

I will draw on my experience working with rodents in Nigeria, including agricultural, household, conservation and public-health contexts. We will also consult people with experience in animal welfare, veterinary science and public health to make sure the questions cover the important issues.

The second goal is to develop a simple but defensible way of assessing the answers. We will consider:

• biological and scientific accuracy;

• animal welfare;

• risks to humans;

• risks to non-target animals and the environment;

• whether the recommendation is practical;

• whether the AI acknowledges uncertainty when it should.

The third goal is to test a number of AI systems using the same questions and compare the types of mistakes they make.

Finally, I want to make the resulting questions and evaluation approach useful to other researchers working on AI and animal welfare. If the pilot works, the same approach could later be extended beyond rodents to other areas of animal use and management.

How will this funding be used?

The funding will mainly support the people needed to do the work properly.

This will include time for developing and testing the questions, obtaining expert reviews, running the AI evaluations, analysing the results and preparing the final report and public materials.

Some funding will also be used for AI model/API access and for meetings with external experts. Because this is a pilot project, I would keep the infrastructure costs relatively low and put most of the funding into research and expert input.

I expect the initial project to run for about 6–9 months.

Who is on your team? What's your track record on similar projects?

I am Olaoluwa John Ademola, PhD, a Lecturer in the Department of Zoology at the University of Ilorin and co-founder of RodentPlus Conservation Initiative in Nigeria.

My academic research has focused mainly on rodents and other small mammals. My PhD research investigated the ecology of Praomys delectorum in the Eastern Arc Mountains of Tanzania. I have also worked on small-mammal surveys and conservation in Nigeria, including research on poorly known montane rodents and shrews.

Through RodentPlus, my work has increasingly moved toward the connection between rodent management, animal welfare and public health. We have worked with farmers and communities on rodent control and have examined the use of rodenticides and other control methods. We are also involved in research into fertility-control approaches as alternatives to conventional rodenticide use.

I work with researchers and practitioners in Africa and internationally, including people working in rodent ecology, animal welfare, veterinary science and rodent fertility control. This existing network will help us bring together the expertise needed to review the AI responses.

I do not have a long track record in AI research, and I would not present myself as an AI safety researcher. The reason I think I can contribute to this area is that I have detailed domain knowledge of rodents and the welfare problems associated with their management. The project is intended to bring that neglected domain knowledge into AI evaluation.

What are the most likely causes and outcomes if this project fails?

The main risk is that the benchmark is too small or too dependent on subjective judgements to tell us much about AI systems.

There may also be disagreement between experts about some answers. Animal welfare questions are not always completely straightforward, and a recommendation that is appropriate in one situation may not be appropriate in another. We will therefore record disagreement rather than pretending that every question has one simple answer.

Another risk is that AI models change during the project. A system that performs poorly today may be updated before the project is completed. We will record the models and versions tested so that the results are easier to interpret.

If the project does not produce a useful benchmark, we should still learn which types of questions are difficult to evaluate and where expert disagreement occurs. If it succeeds, we will have a practical way of testing an area of AI behaviour that is currently receiving relatively little attention.

How much money have you raised in the last 12 months, and from where?

RodentPlus and my related research activities have received project support during the past year from a number of conservation and animal-welfare funders, including Animal Charity Evaluators, Indianapolis Zoo and ZGAP.

Most of this funding is restricted to particular projects, such as conservation research, animal welfare, equipment or community outreach. It is therefore not available to fund this proposed AI project.

I will provide the exact total and individual amounts in the final application based on our financial records.